An Asymmetric Contrastive Loss for Handling Imbalanced Datasets

Area of Science:

• Machine Learning
• Computer Vision
• Data Science

Background:

• Contrastive learning (CL) is a powerful representation learning technique that excels at grouping similar samples in feature space.
• While CL is effective for imbalanced datasets, existing methods lack specific modifications to handle class underrepresentation.
• Addressing class imbalance is crucial for robust performance in real-world machine learning applications.

Purpose of the Study:

• To introduce an asymmetric version of contrastive learning (ACL) specifically designed to tackle class imbalance.
• To propose asymmetric focal contrastive loss (AFCL) as an advancement over ACL and focal contrastive loss (FCL).
• To evaluate the efficacy of AFCL in improving classification accuracy on imbalanced datasets.

Main Methods:

• Developed an asymmetric contrastive learning (ACL) framework.
• Introduced asymmetric focal contrastive loss (AFCL), generalizing ACL and focal contrastive loss (FCL).
• Conducted experiments on imbalanced Fashion MNIST (FMNIST) and ISIC 2018 datasets.

Main Results:

• The proposed asymmetric focal contrastive loss (AFCL) demonstrated superior performance compared to standard contrastive loss (CL) and focal contrastive loss (FCL).
• AFCL achieved higher weighted and unweighted classification accuracies on imbalanced datasets.
• The asymmetric approach effectively mitigates the negative impact of class imbalance on representation learning.

Conclusions:

• Asymmetric contrastive learning (ACL) and asymmetric focal contrastive loss (AFCL) offer effective solutions for representation learning on imbalanced datasets.
• AFCL provides a significant improvement over existing contrastive learning methods in scenarios with class imbalance.
• This work contributes a novel approach to enhance the robustness and accuracy of machine learning models facing data imbalance.